From: Transformer-CNN: Swiss knife for QSAR modeling and interpretation
Layer | Relevance, R (L + 1) | Relevance, R (L) | Delta, R (L + 1)-R (L) | Bias, Delta / R (L + 1) *100% |
---|---|---|---|---|
Result | 0.98119 | – | – | – |
HighWay Output | 0.98119 | 0.9300 | 0.0512 | 5.21 |
HighWay Input | 0.9300 | 0.7227 | 0.2073 | 22.3 |
DeMaxPool | 0.7227 | 0.7371 |  − 0.0144 |  − 1.98 |
Conv1 | 0.0090 | 0.0117 |  − 0.00271 |  − 30.1 |
Conv2 | 0.1627 | 0.1627 | 0a | 0 |
Conv3 | − 0.0443 |  − 0.0443 | 0 | 0 |
Conv4 | 0.0191 | 0.0191 | 0 | 0 |
Conv5 | − 0.0984 |  − 0.0984 | 0 | 0 |
Conv6 | − 0.0136 |  − 0.0136 | 0 | 0 |
Conv7 | 0.0806 | 0.0806 | 0 | 0 |
Conv8 | 0.0957 | 0.0957 | 0 | 0 |
Conv9 | 0.1528 | 0.1528 | 0 | 0 |
Conv10 | 0.0845 | 0.0845 | 0 | 0 |
Conv15 | 0.1038 | 0.1038 | 0 | 0 |
Conv20 | 0.1851 | 0.1851 | 0 | 0 |
Total | 0.98119 | 0.7398 | 0.2414 | 24.6 |